Comparing two different controller designs for evolving robots

1
Comparing two different controller designs for
evolving robots

• Ferenc Havasi
• Vincenzo Giordano
• Michael Schwarz
• Gregory Valigiani
• Stefan Wiegand
• Thanks to Marc Schoenauer and Nicolas Godzik

2
Evolutionary robotics

• The control system of the robot evolves through
  ES
• The fitness function is set according to the task
  to be performed
• Explorative evolution, minimal human intervention

3
Advantages

• self-organization, no explicit design
• Exploitation of the interactions between the
  robot and the environment
• No decomposition in and integration of
  (hopefully) simpler tasks
• No need of many assumptions (not knowledge-rich
  representation)

4

• Disadvantages
• The automatic process of adaptation on
  real-robots is time-consuming
• Need for heuristic and rules of thumb (trial and
  error procedures)
• Lack of formal principles Its difficult to
  compare slightly different experiments

5
Main point

• What happens if we add more knowledge in the
  evolution process?
• Are we reducing capability of evolution of
  exploring?
• Are we reducing capability of evolution of
  generalizing?

6
Two different architectures
7
Experimental set-up

• Tasks
• Structure of the neural network (try simple ones
  before)
• Structure of the evolutionary strategy
  (20,140)-ES with weak elitism
• Fitness function (as simple as possible)

8
Experiments

• Obstacle avoidance
• fitness high speedstraight motionlow sensor
  activation
• Following a moving object
• same fitness
• but
• if the camera is on then slow down and go
  straight

9
Obstacle avoidance
10
Following the object
11
Generalization

• We want to measure the ability to generalize the
  acquired knowledge to novel circumstances
• We evaluate the last generation of individual for
  each run in a different environment

12
Generalization (obstacle avoidance)
13
Generalization (following object)
14
Conclusions

• The following object experiment was successfull
  with both controllers
• The second controller seems to perform better and
  to evolve faster than the first one
• More specific design doesnt necessarily lead to
  better results
• Its necessary to test generalization
  capabilities performing different tasks